Microsoft: Phi Series and Copilot

Learning Objectives

• Compare Microsoft's dual role as an OpenAI partner and its own model developer through the Phi family
• Explain the "textbook quality data" hypothesis behind Microsoft's Phi series and its implications
• Describe what Microsoft 365 Copilot does and why it is the most pervasive enterprise AI deployment globally

Microsoft: Copilot and the Phi Series

Microsoft occupies an unusual position in the AI landscape. Through its $13 billion+ investment in OpenAI it has been the primary cloud partner for OpenAI's models via Azure OpenAI Service. The exclusivity was loosened by the April 2026 amended agreement (see "The OpenAI Partnership" section below), but Microsoft remains OpenAI's largest single shareholder and dominant commercial partner. In parallel, Microsoft also develops its own models — primarily the Phi series — optimized for efficiency and on-device deployment.

Phi-4: Small Model, Large Capability

Phi-4 is Microsoft's research model demonstrating that small, carefully trained models can match or exceed much larger models on specific capabilities.

Key characteristics:

• MIT license — fully open source, including commercial use
• Exceptional reasoning at small size — outperforms much larger models on mathematics and coding benchmarks
• On-device capable: runs on laptops and modern smartphones
• 16K–32K context window (smaller than frontier models, but appropriate for on-device use cases)
• Available through Hugging Face and Azure AI Foundry

The "textbook quality data" hypothesis: Microsoft's Phi series research demonstrates that training data quality can substitute for parameter scale. Phi models are trained on carefully curated "textbook quality" content — mathematical problem sets, high-quality code, structured explanations — rather than the raw internet text used for larger models.

Phi-4 variants extend the family's reach:

• Phi-4-multimodal (5.6 billion): Speech, vision, and text in a single small model — designed for multimodal on-device applications
• Phi-4-mini (3.8 billion): Even more compact for the most constrained deployment environments
• Phi-4-reasoning (14 billion): Larger variant optimized for complex reasoning tasks while remaining small enough for single-GPU deployment

Phi-4

Microsoft

Open Source

Strengths

MIT license; exceptional reasoning/math at small size; on-device capable; 16K–32K context

Context Window

32K tokens

Pricing

Free (MIT license)

Microsoft →

Microsoft 365 Copilot: Enterprise AI at Scale

Microsoft 365 Copilot is the most deployed enterprise AI product in the world — AI embedded across the entire Microsoft Office suite used by hundreds of millions of knowledge workers.

What it includes:

• AI in Word: Draft, rewrite, summarize documents
• AI in Excel: Generate formulas, analyze data, create charts from natural language
• AI in PowerPoint: Generate presentation slides from documents or outlines
• AI in Outlook: Draft emails, summarize threads, schedule meetings
• AI in Teams: Real-time meeting transcription, summaries, action items
• Multi-agent orchestration: Agents across M365 apps can collaborate — an agent in Teams can pull data from Excel, draft content in Word, and schedule a meeting in Outlook

Pricing: $30/user/month enterprise add-on to M365 subscriptions.

Copilot Wave 3 and Leadership Changes (March 2026)

M365 Copilot Wave 3 (announced March 9, 2026) introduces multi-model intelligence — Copilot now uses models from multiple providers, including Anthropic's Claude alongside OpenAI's GPT models. The new "Copilot Cowork" feature enables agents to collaborate across M365 apps with richer context.

In a significant leadership shakeup on March 17, Mustafa Suleyman (former Google DeepMind co-founder) shifted from leading Copilot to heading Microsoft's "Superintelligence" efforts. Jacob Andreou (former Snap executive) was named EVP of Copilot, reporting directly to CEO Satya Nadella.

Microsoft also rolled back some Copilot AI integrations from Windows 11 (March 20, 2026) — reducing AI features in Photos, Widgets, Notepad, and Snipping Tool — a pragmatic acknowledgment that not every AI integration improves the user experience.

Pricing changes effective July 1, 2026: M365 E3 rises from $36 to $39/user/month; E5 from $57 to $60. The $30/user/month Copilot add-on price remains unchanged.

The OpenAI Partnership: From Exclusivity to Primary Partner (April 2026)

Microsoft's $13 billion+ OpenAI investment has become one of the most valuable corporate bets in history. After OpenAI's October 2025 recapitalization, Microsoft's equity stake is valued at approximately $135 billion (~27% diluted). Microsoft earned $7.6 billion from OpenAI in a single quarter (January 2026 earnings).

On April 27, 2026, Microsoft and OpenAI announced an amended agreement that materially restructures the seven-year partnership:

• Cloud exclusivity ends. Microsoft's exclusive IP license becomes non-exclusive through 2032. OpenAI can now ship products on any cloud (AWS, Google Cloud, Oracle, or its own infrastructure). Microsoft remains the primary cloud partner with first-shipping rights when it has capacity to support new models.
• Revenue-share reversed. Microsoft stops paying revenue share to OpenAI. OpenAI continues paying Microsoft a capped revenue share through 2030.
• $250 billion Azure commitment. OpenAI contractually commits to spending $250 billion on Microsoft Azure services, ensuring the partnership remains commercially central even as exclusivity formally ends.
• Equity preserved. Microsoft keeps its ~27% stake — making the upside-on-OpenAI thesis intact even as the day-to-day commercial structure becomes less restrictive.

The amendment also explains why Copilot Wave 3 (March 2026) was able to introduce multi-model intelligence — drawing on Anthropic's Claude alongside OpenAI's GPT — and signals a structural shift: OpenAI moves from a Microsoft-bound subsidiary-style relationship toward a more independent strategic partner, while Microsoft retains the equity upside, the Azure revenue, and primary-partner status without locking customers (or itself) into a single-model future.

Key Takeaways

• Microsoft's Phi-4 family (including multimodal, mini, and reasoning variants) demonstrates that training data quality can substitute for parameter scale — MIT license makes it accessible for on-device and edge use
• Phi models run on consumer hardware (laptops, smartphones), enabling on-device AI for privacy-sensitive applications
• Microsoft 365 Copilot Wave 3 now uses multi-model intelligence (Anthropic's Claude alongside GPT) — structurally enabled by the April 2026 OpenAI agreement amendment that ended cloud exclusivity
• Microsoft's $13 billion+ OpenAI investment is now worth ~$135 billion (~27% equity); Microsoft earned $7.6 billion from OpenAI in a single quarter; OpenAI contracted $250 billion in Azure services through the new amendment
• Leadership shakeup (March 2026): Mustafa Suleyman moved to superintelligence research; Jacob Andreou now leads Copilot